Research And Application Of Fusion Matching Network And Prototype Network In Few-Shot Relation Extraction

With the rapid development of large data and the Internet,traditional structured data,because of its too abstract characteristics,has a decreasing proportion in people’s daily life.In contrast,unstructured data,such as pictures,text,video,audio and other data forms,accounts for the vast majority.This shift also presents a new challenge,how to extract information from large amounts of unstructured data.Relationship extraction is one of the key tasks in information extraction.It learns the semantic information and entity information in text data,and uses the learned features to batch automate the relationship extraction of large amounts of unstructured data,such as text information,to support the practical application of knowledge base,question and answer system,information retrieval and so on.However,at present,most of the information extraction is based on deep learning methods.Although the effect is significant,deep learning models often require a large amount of high-quality labeled data,which is too costly in the actual application scenarios.How to make the model learn with a small number of samples and get a better relationship extraction effect is a very meaningful research direction.Therefore,this paper takes the maternal and infant product review as an example,and achieves the extraction of small sample relationships in the Chinese context through language model with matching network,prototype network and other modules.The main contents of this paper are as follows.(1)Method based on matching network.In the current algorithm research,the vast majority rely on the English context,but less on the Chinese context.At the same time,the grammar,the morphological features and the pronunciation features of Chinese have more features than those of English.Therefore,a Chinese language model combining characters,glyphs and pronunciation is proposed.Considering the outstanding performance of hint learning in the small sample area,this paper designs a multi-tailed pointer method to stitch the hint templates after the corpus,and complete the relationship extraction through the identifier,which has a certain improvement over other similar methods.(2)Method based on prototype network.Prototype network has always been one of the commonly used methods in the small sample problem.The idea of prototyping is especially used to model the relationship in the relationship extraction problem.Therefore,this paper constructs a small sample relationship extraction from another perspective by combining the text representation ability of language model,the feature of entity location information,the semantic information of corpus,and the method of contrastive learning and KL dispersion.Compared with similar methods,it also achieved some improvement.(3)Integration framework of few-shot relation extraction.This paper integrates the hint-based learning method and the prototype network-based method by averaging the class probability to achieve a certain degree of complementarity.(4)Application of few-shot relation extraction.Although the few-shot relation extraction model is not as accurate as the relationship extraction model trained by a large amount of data at present,it requires less labor and time costs.Therefore,this paper designs an intelligent labeling system based on small sample relationship extraction and a method to rapidly build domain knowledge map.